cold-read-gate

# Cold-Read Gate Text written inside a long session inherits the session's context — codenames, version baselines, internal PR numbers, "the obvious fix" — and the author can no longer see which parts won't survive contact with a reader who has none of it. Before publishing an outward-bound artifact, **dispatch an isolated, cheap, context-free agent to read it cold and interrogate it**. What confuses the cold reader is what will cost you a round-trip question (or a silent deprioritization) from the real audience. The reader is deliberately **haiku, isolated, and shown only the artifact**. This is a measured exception to the "always Opus for subagents" rule: the subagent here is not doing delegated work — it *is* the measurement instrument. The test is "can a low-context reader act on this text alone?", and a stronger model (or one with session context) would answer a different, easier question. If haiku can act on it, a busy maintainer can. ## When to Use This Skill "Outward-bound" means any audience that lacks your session context — external maintainers *and* future teammates reading internal docs cold both qualify. | Use this skill when... | Skip when... | |---|---| | Filing issues/MRs on an upstream tracker (external maintainers) | Internal scratch notes, commit messages, chat replies | | Publishing docs a new team member will land on cold | The artifact is throwaway or has a captive expert audience | | PR descriptions for reviewers outside the work's context | The te...